littlewhitesea / STDAN-CVPRW-2022

Space-Time Video Super-Resolution Using Deformable Attention Network

STDAN-CVPRW-2022 (Keep Update)

This is the official Pytorch implementation of Space-Time Video Super-Resolution Using Deformable Attention Network.

Dependencies

• Python 3.8.0 (Recommend to use Anaconda)
• PyTorch == 1.8.0
• CUDA == 10.1
• Deformable Convolution v2, we use Detectron2's implementation in the network.

Installation

1. Install the required packages:

   pip install -r requirements.txt
   

2. Compile the deformable attention and convolution:

   cd YOUR_PATH/STDAN/codes/models/modules/DCNv2_latest
   bash make.sh
   

Training

Dataset preparation

You require to prepare datasets for training the model, the detailed information is elaborated in Data Preparation.

Train the STDAN model

cd YOUR_PATH/codes
python train.py -opt options/train/train_stdan.yml

Testing

Pretrained Models

Our pretrained model can be downloaded via Google Drive or Baidu Netdisk(access code: gnhk). After you obtain the pretrained model, please put them into the YOUR_PATH/experiments/pretrained_models folder.

Testing on Vid4/SPMC datasets

cd YOUR_PATH/codes
python test.py

Testing on Vimeo-Slow/Medium/Fast datasets

cd YOUR_PATH/codes
python test_vimeo.py

Acknowledgments

We thank Zooming Slow-Mo and Detectron2. They provide many useful codes which facilitate our work.